Read the AI Summary and You May Remember Its Mistake

An AI summary that gets one detail wrong can change what you remember. In a 2026 study, people watched a short accident video and later read a summary naming the wrong traffic sign. Only 44.8% then recalled the sign correctly, against 83.6% after an accurate summary.¹ Believing a human wrote it did not help.

Mattea Sim (Georgetown University), Yael Eiger (University of Washington) and Tadayoshi Kohno (Georgetown) posted AI-Enabled Human Memory Manipulation to arXiv on 23 September 2026. A shortened version appears in the proceedings of AIES 2026, the AAAI/ACM conference on AI, ethics and society.¹² Their abstract names the assumption the result breaks:

Though "humans-in-the-loop" are often expected to correct for AI's mistakes, our work suggests human memory can instead be distorted by these mistakes.¹

The study used animated car-crash clips, not meetings. Everything below about meeting recaps is an inference from that design, and it is labeled where it appears. An earlier post covered the storage half of this problem — keep the original beside the summary. This one is about the reader.

The study, briefly: a 25-second crash and a summary read a day later

Participants watched one of two 25-second animated videos of a car hitting a pedestrian, identical except for the sign at the intersection: stop or yield. Between 24 and 48 hours later they read a 21-sentence ChatGPT summary of the event, then answered twelve two-choice memory questions. The critical one asked which sign the car had stopped at.

The design is a conceptual replication of a 1978 experiment by Elizabeth Loftus and colleagues. Their 1,242 subjects saw slides of "a single auto-pedestrian accident" and were then given consistent, misleading or irrelevant information about it.³ The 2026 version delivers the detail through AI-written text, and 328 adults recruited online made it into the analysis.¹

Before reading, each participant was told the summary came from an AI model or from a person. That was a cover story — everyone read the same ChatGPT text. Participants also completed a ten-item trust-in-AI scale and reported how often they use chatbots.¹

One detail matters for reading the result fairly. The researchers took one ChatGPT summary of the yield-sign video, removed the details that differed between the two clips, and set the sign to stop or yield. For about half the readers, it named the sign they had not seen.¹

The summary got one detail wrong because the experiment put it there — the right control for isolating what a wrong detail does to a reader.

What they actually measured: one wrong sign, recall nearly halved

After a summary that matched the video, 83.6% of participants named the correct sign. After a summary that named the other sign, 44.8% did — on a two-choice question, worse than a coin toss. Prior work with no summary found 80.26% correct, so the accurate summary seems to have added little and the wrong one took much away.¹

The damage was targeted. The same people remembered the rest of the clip well: 96.6% correctly said the accident happened in daytime, and across 11 non-critical questions they answered 91.4% correctly.¹ Their memory of the scene held — the detail that moved was the one the summary contradicted.

The label made no difference. Among readers of the misleading summary, 46.3% answered correctly when told an AI wrote it and 43.4% when told a person had.¹ Trust in AI made no difference either. The paper reports "no significant interactions", "including the central three-way interaction" (p = .820), and the same null for how often people use chatbots (p = .547).¹

Skepticism toward AI — the instinct most people assume protects them — was not a filter. Sim put the consequence plainly in Georgetown's announcement: "AI is a new method of delivering misinformation, and it has the potential to create these false memories for people who are reading that information".²

The summaries also left out the crash

The study's other half asked how often AI summaries get an event wrong with no one's help. ChatGPT-5.5 and Gemini 2.5 Flash-Lite each summarized both videos five times. Across those 20 summaries, 51.6% of central details were omitted on average, and all but one left out the collision itself — the event the video exists to show.¹

The paper's wording is blunt: "all but one summary (95% of summaries) omitted the single most central detail of the event: the car colliding with the pedestrian."¹ Ninety percent also dropped the pedestrian lying on the ground and getting up. One summary went further and reported that "no other characters or vehicles interacted directly with the scene."¹

Omission dominated. Outright inaccuracies touched 6.1% of central details, against 51.6% omitted.¹ Still, "All twenty summaries included mistakes", with at least seven errors each, "notably high in summaries that averaged just 305 words."¹

Yael Eiger said in the same announcement: "I was struck by how bad the summaries were, even at this stage in AI development".² The summary used in the memory experiment had been re-prompted until it included the crash.¹ Real recaps get no such curation.

Omission is one failure; overstating is another, covered in your AI summary says more than your notes do.

Why this matters for the meeting recap you read the next morning

No meeting was tested. The inference is short, though: an AI recap is a summary you read after an event you attended, often a day later, which is close to the study's design. If a recap drops or changes a decision, this result suggests your memory may drift toward the recap rather than catch the error.

Read the AI summary of a meeting you attended, in other words, and you may remember its mistake.

Georgetown's release names the setting. AI summaries are spreading "from news article previews to workplace meeting transcriptions to high-stakes settings like police body camera footage".² Naming a setting is not testing it — the paper never ran a meeting.

Vendors place the check on you. Microsoft's Teams help page, under the Copilot "Recap the meeting" prompt, says: "Be sure to verify your results because AI-generated content could be incorrect."⁴ Verification assumes a clean reference — and once you have read the recap, your memory may no longer be one.

A Hacker News commenter who runs several teams described the timing problem in April 2025: "Every mistake the AI makes is completely understandable, but it's only understandable because I was in the meeting and I am reviewing the notes right after the meeting. A week later, I wouldn't remember it, which is why I still just take my own notes in meetings."⁵

Checking is hard even before memory enters. In a CHI 2024 study of ChatGPT answers to Stack Overflow programming questions, "52% of ChatGPT answers contain incorrect information", and participants "overlooked the misinformation in the ChatGPT answers 39% of the time".⁶

Timing may matter too. Loftus's 1978 team found "misleading information had a larger impact if introduced just prior to a final test rather than immediately after the initial event."³ Carried over to meetings — an inference, not a finding — a recap skimmed as the only refresher before the follow-up is the worst-placed version.

Writing it down first is not a vaccine

The obvious fix — recall the meeting yourself before reading the recap — has a complicated record. In eyewitness studies, recalling an event first sometimes made people more susceptible to later misinformation, and a written account did not reduce susceptibility in the newest test. Two things did remove the extra effect: a warning, and checking for discrepancies.

Jason Chan, Ayanna Thomas and John Bulevich expected recall to protect witnesses, since testing usually strengthens memory. Their 2009 result went the other way: "We show, however, that immediate cued recall actually exacerbates the later misinformation effect for both younger and older adults."⁷ The literature now calls this retrieval-enhanced suggestibility, or RES.

Emily Bird, Jan Wiener and Janice Attard-Johnson tested the Self-Administered Interview, a written recall form built for witnesses, against plain written free recall. With 141 participants exposed to misinformation after a delay: "Surprisingly, neither the SAI nor the WFR reduced susceptibility to misinformation."⁸

Writing still helped in other ways — the interview form improved immediate recall, and completing it reduced incorrect answers later compared with no immediate recall. Earlier studies of the same form also recorded fewer misinformation items.⁸ The evidence is mixed, not negative.

What removed the extra susceptibility was simpler. In a 2010 study, subjects heard a recap of a video and were told before the final test that its accuracy could not be verified. "However, this RES effect was eliminated when subjects were warned about the veracity of the narrative."⁹ The Sim paper adds the counterweight: "warnings are not always effective".¹

B. J. Butler and Elizabeth Loftus found a second lever inside reading itself: "In Experiments 2 and 3, having subjects concurrently assess the consistency of the misinformation narratives negated the RES effect."¹⁰ Prior recall raised susceptibility "only for those who failed to detect discrepancies".¹⁰

Note the scope — both remedies removed the extra effect that recalling first created. Neither made anyone immune.

How to apply this tomorrow: five habits, labeled honestly

The evidence supports two habits directly: expect the recap to be wrong, and check it while you read. The other three are inference — they give you a record the recap cannot touch. None of them shields memory. They give you something besides memory to check against, which is what the lab results point toward.

1. Write the decisions down before you open the recap. Three to five lines: what was decided, who owns it, by when. Its value is as a separate dated record, not as protection for your memory — the Chan and Bird results rule out that promise. It is the same logic as writing a prediction down before hindsight edits it. (Inference.)
2. Open the recap expecting errors. The 2010 subjects were told, after the recap and before the test, that its accuracy could not be verified. (Evidenced in the lab; not always effective.)
3. Read it as a checker. Mark each line that differs from your own record as you go. Checking for discrepancies while reading is what negated the effect in Butler and Loftus. (Evidenced in the lab.)
4. Look for the central decision first. In the study, 95% of summaries left out the collision. Absence is the likeliest error, and the hardest to notice. (Inference.)
5. Settle disagreements against the recording or transcript, never memory. Keep the source linked from the note; the convention is in keep the original beside the summary. (Inference.)

The order matters more than any single step — your record first, the recap second, the source last.

The caveats

The finding is strong and the setting is narrow. Two 25-second animated clips are not a ninety-minute planning meeting, one traffic sign is not a list of decisions, and a lab recap carries no stakes. The remedy studies used videos, slides and narratives, not workplace summaries. Read the advice above as reasoning built on them.

• Videos, not meetings. The authors say so: "This analysis was based on two videos previously utilized in the memory literature depicting video game footage of a simple car accident, which limits the generalizability of this work to other more realistic videos."¹
• The error was placed, not generated. The misleading condition was a summary that named the other video's sign — a planted error, not an AI one. The omission figures show real summaries do miss central details, but the 44.8% does not measure an AI's own mistake.
• Recognition, not free recall. Participants picked one of two signs. Remembering a decision unprompted is a different task.
• One delay. The summary came 24 to 48 hours after the video, and memory was tested the same session. Longer effects were not measured.
• A baseline borrowed from prior work. The 80.26% no-summary figure comes from an earlier study the paper cites, not from this sample.¹
• Two versions of the paper. The numbers here come from the full arXiv version; the shortened AIES proceedings version reports the same headline figures.

Frequently Asked Questions

The questions below are phrased the way people search for them. The short version: a wrong detail in a summary moved memory in the lab, labels and scepticism did not stop it, and the fixes with evidence behind them are expecting errors and checking while reading.

Can AI meeting summaries change what I remember?

Possibly. No study has tested meetings, but a 2026 experiment with 328 adults found that reading a ChatGPT summary edited to name the wrong traffic sign cut correct recall of the sign from 83.6% to 44.8%. A meeting recap read a day later follows a similar pattern, so the risk is a reasonable inference rather than a proven one.

Are AI meeting notes reliable?

Not reliably enough to replace a check. When researchers asked ChatGPT and Gemini to summarize two short accident videos, every one of 20 summaries contained mistakes, and 95% omitted the collision itself. Microsoft's own Teams help page tells users to verify Copilot meeting recaps because AI-generated content could be incorrect.

Should I write my notes before reading the AI recap?

Yes, but for the right reason. A few lines on what was decided give you a separate record to compare against. They are not proven to protect your memory: in eyewitness studies, recalling first sometimes raised susceptibility to misinformation, and a written account did not reduce it in a 2024 test.

Does knowing a summary was written by AI protect me?

No, according to the 2026 study. Readers of the misleading summary answered correctly 46.3% of the time when told it came from AI and 43.4% when told a person wrote it. How much people trusted AI, and how often they used chatbots, made no significant difference to how far they were misled.

Can I check an AI summary against my own memory?

Only before you read it. Once a summary has been read, its errors can blend into what you recall, so memory is no longer an independent reference. Check the summary against something it cannot change: your own notes written beforehand, the recording or the transcript. Settle disagreements against the source.

What reduces the effect of a misleading summary?

Two things removed the extra susceptibility in lab studies: being warned before the test that a recap's accuracy could not be verified, and checking the recap for discrepancies while reading it. Neither made people immune, and the study's authors note that warnings are not always effective. Treat both as habits, not guarantees.

A recap you have read becomes part of what you remember; a note written before it is the one witness the recap cannot reach.

mnmnote.com keeps meeting notes as plain Markdown on your own device, so the lines you wrote before the recap stay the lines you wrote.